Shape optimization and active flow control for improved aerodynamic properties Siniša Krajnovic

Transcription

1 Shape optimization and active flow control for improved aerodynamic properties Siniša Krajnovic

2 HPC resources used Computer resources at C3SE at Chalmers in Göteborg Computer cluster: Neolith NSC Linköping 6440 cores

3 Automatic aerodynamical Shape Optimization (Students E. Helgason and H. Hafsteinsson) Programs used: Fire Sculptor modefrontier Car model used: Full scale experimental model from Volvo Cars named the VRAK

4 Process AVL/Fire: Mesh generation Sculptor: Create volume for mesh morphing modefrontier: Adjust control parameters for mesh morphing Sculptor: Mesh morphing AVL/Fire: CFD calculations modefrontier: Collects results and change mesh morphing parameters modefrontier: Optimal solution selected

5 Control volume set in Sculptor Morphing the rear end of the car

6 Workflow in modefrontier Here one input variable (Rear_r) controls the mesh deformation in Sculptor ModeFrontier adjusts the control variable and collects results for Cd Built in optimization algorithms in modefrontier can be used, i.e. SIMPLEX or the gradient based algorithm NLPQLP

7 Deformed mesh Upper right fig: Upper limit of the deformation parameter (Rear_r = 0.3) Lower left fig: Undeformed car (Rear_r = 0) Lower right fig: Lower limit of the deformation Paramater (Rear_r = -0.2)

8 Results from modefrontier Automatic Optimization with SIMPLEX algorithm using steady k-e turbulence model with inlet velocity of 10 m/s. Using course mesh with approx cells Each simulation runs from t=0. The control variable Rear_r = 0, corresponds to the original car

9 Flow visualization Chalmers University of Technology U = 10 m/s

10 Automatic Optimization with the NLPQLP algorithm using unsteady k-z-f turbulence model with inlet velocity of 140 km/h Using finer mesh with boundary layers, approx cells First run is the original VRAK Small modifications are made on the surface and the simulation is restarted with results from previous simulation Each modification runs for 0.5s Last 0.1s gives average Cd Horizontal dotted line represents experimental value of the drag coefficient for orginal VRAK Cd-exp = 3.05 Vertical lines emphasize at what time simulation is restarted with new deformed geometry

11 Results from ModeFRONTIER Cd is reduced by 3% in four simulations.

12 Modifying the car surface in Sculptor Upper right fig: Upper limit of the Deformation parameter, Rear_r = 0.1 Lower left fig: Undeformed car, Rear_r = 0.0 Lower right fig: Lower limit of the deformation Paramater, Rear_r = -0.1

13 Deformed mesh Upper right fig: Upper limit of the Deformation parameter, Rear_r = 0.1 Lower left fig: Undeformed car, Rear_r = 0.0 Lower right fig: Lower limit of the deformation Paramater, Rear_r = -0.1

14 Workflow in ModeFrontier modefrontier adjusts the control variable and collect results for Cd from Fire Built in optimization algorithm can be used to minimize Cd ES is chosen based on how many concurrent designs it can run modefrontier and Sculptor run locally One deformation is performed at a time The mesh is transferred to the cluster CFD calculations are restarted using the new mesh Each design takes ~ 22 h All flow results for each design can be obtained from the cluster

15 System specs Computer cluster: Neolith NSC Linköping 6440 cores Processor: Intel Xeon E5345 Quad Core Processor 2.33 GHz, 4MB Level cache Interconnect: Infiniband ConnectX interconnect Node memory: 16 GiB Computer resources at C3SE at Chalmers in Göteborg Number of cells Simulation runs on 48 CPUs Time step execution time ~ 80s Time step T = 0.001s Time for simulation to run 1.0s ~ 22h A particle will pass the car 10 times during 1.0s

16 Results from modefrontier Cd is decreased by 8.8% Four concurrent simulations are made each time. 8 DOE points are equally distributed over the design space. Optimization algorithm (ES) is used locally around the best point found in the DOE sequence.

17

18 Introduction Task Minimize rolling and yawing moments of a train Programs AVL FIRE Mesh creation and CFD simulations Sculptor Mesh deformation modefrontier - Optimization

19 The Optimization Process Mesh Generation Optimization Mesh Deformation CFD Simulation

20 Computational Domain Chalmers University of Technology

21 Computational Domain 30 side wind U = 30 m/s

22 Mesh deformation in Sculptor Creation of ASD volume

23 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = = 0.000

24 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = = 0.000

25 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = = 0.000

26 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = = 0.000

27 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = = 0.000

28 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = =

29 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = = 0.000

30 Deforming the train surface in Sculptor Deformation parameter 1 [-0.002,0.004] 2 [-0.004,0.004] 1 = = 0.004

31 Deforming the train surface in Sculptor Deformation parameter 2 2 [-0.004,0.004] A) 1 = B) 1 = C) 1 = = = = = A) 1 = = = = B) C)

32 Original 1 [-0.002,0.004] 2 [-0.004,0.004]

33 Optimization Turb. Model steady k-z-f U [m/s] Num. Cells Deform. Par. Objective Opt. Alg min M r, M y ES

34 Workflow in modefrontier Two input variables, 1 and 2 Two objectives, minimize M r and M y Optimization algorithm, Evolution Strategy (ES) modefrontier and Sculptor run locally AVL FIRE runs on cluster Each design is restarted from original train DOE Points Concurrent Designs Size of Generation Generations Simulation Time [ h ] CPU s Total CPU Time [ h ]

35 Results 1 2 My [Nm] % Mr [Nm] % Original DOE

36 Results 1 2 My [Nm] % Mr [Nm] % Original DOE ES

37 Original Optimized

38 Original Chalmers University of Technology

39 Optimized Chalmers University of Technology

40 The active flow control problem 1. Reference experimental work [1] Henning et al. Feedback control applied to the bluff-body wake. In King R. (ed.), Active Flow Control, Springer-Verlag, Open and closed-loop control; Re h in the range of ; Harmonic actuation in time through two spanwise slots at 45 o with the streamwise direction; Drag reduction of 15% at St A =0.17, in-phase actuation.

41 3. Boundary conditions: Model and computational details At the slots, oscillatory forcing is implemented as: The actuation amplitude follows from momentum coefficient:

42 Model and computational details 4. Resolution and numerical details: Total number of nodes ; Spatial resolution acording to Physical time step = ; 96.5% of the cells with CFL < 1; Space discretization: 2nd order central differences; Temporal discretization: Three-time-level Scheme (implicit second order scheme); Solution algorithm: SIMPLE; Turbulence model: LES Smagorinsky Model; C s =0.1.

43 Drag control results Drag control results from the LES at Re h =2 10 4, St A =0.17 and C μ =0.015: 11% drag reduction achieved; 20% pressure recovery in the near-wake region;

44 Exploring the flow The time-averaged flow: Natural flow Controlled flow Reduction of the thickness of the upper and lower edge thin vortices; Foci C 1 /C 2 and the saddle point are displaced further downstream by 20% from their streamwise locations in the natural case.

45 Comparison of time-averaged flows Controlled flow Natural flow

46

47 How the communication beetween Matlab and Fire works

48 Phase control Chalmers University of Technology